Re: patch: linux-2.6.4/lib/int_sqrt.c long aware

Re: patch: linux-2.6.4/lib/int_sqrt.c long aware

[Randy.Dunlap]

| From: <> (Walter Harms)
| 
| hi list,
| this patch make int_sqrt() handle long properly. i tested it
| on my alpha and i386. also some comparision again an alternativ int_sqrt() code.
| The placement (and using) of 'tmp' is a bit odd but faster (on alpha, no matter on i386).
| 
| happy hacking,
| walter
| 
| ps: i must use c&p so be aware of possible tab->space translations
| 
| 
| --- linux-2.6.4/lib/int_sqrt.c.org	2004-03-21 12:58:55.783035928 +0100
| +++ linux-2.6.4/lib/int_sqrt.c	2004-03-21 13:16:20.999138928 +0100
| @@ -10,22 +10,23 @@
|   */
|  unsigned long int_sqrt(unsigned long x)
|  {
| -	unsigned long op, res, one;
| +	unsigned long op, res, one,tmp;
|  
|  	op = x;
|  	res = 0;
|  
| -	one = 1 << 30;
| +	one = 1UL << (BITS_PER_LONG-2);
|  	while (one > op)
|  		one >>= 2;
|  
|  	while (one != 0) {
| +		tmp=res + one;
|  		if (op >= res + one) {
| -			op = op - (res + one);
| -			res = res +  2 * one;
| +			op = op - tmp;
| +			res = tmp + one;
|  		}
| -		res /= 2;
| -		one /= 4;
| +		res >>= 1;
| +		one >>= 2;
|  	}
|  	return res;
|  }

Hi Walter,

I found a little time to play with this.  I only tested it on
a P4 UP machine (via userspace app).
It does indeed speed up integer sqrt typically by around 3% for me.

--
~Randy

patch: linux-2.6.4/lib/int_sqrt.c long aware

[Walter Harms]

hi list,
this patch make int_sqrt() handle long properly. i tested it
on my alpha and i386. also some comparision again an alternativ int_sqrt() code.
The placement (and using) of 'tmp' is a bit odd but faster (on alpha, no matter on i386).

happy hacking,
walter

ps: i must use c&p so be aware of possible tab->space translations


--- linux-2.6.4/lib/int_sqrt.c.org	2004-03-21 12:58:55.783035928 +0100
+++ linux-2.6.4/lib/int_sqrt.c	2004-03-21 13:16:20.999138928 +0100
@@ -10,22 +10,23 @@
  */
 unsigned long int_sqrt(unsigned long x)
 {
-	unsigned long op, res, one;
+	unsigned long op, res, one,tmp;
 
 	op = x;
 	res = 0;
 
-	one = 1 << 30;
+	one = 1UL << (BITS_PER_LONG-2);
 	while (one > op)
 		one >>= 2;
 
 	while (one != 0) {
+		tmp=res + one;
 		if (op >= res + one) {
-			op = op - (res + one);
-			res = res +  2 * one;
+			op = op - tmp;
+			res = tmp + one;
 		}
-		res /= 2;
-		one /= 4;
+		res >>= 1;
+		one >>= 2;
 	}
 	return res;
 }